The weights are not released, and this model is not served on Logits today. This repository holds its model card and results, not the model.
GLM-5.3 with its weights edited along the refusal direction by Logits. On JailbreakBench it refuses 1% of harmful requests in English and 11% in Chinese, against 93% and 90% for the original. It keeps its knowledge scores and probably loses about 3 points on IFBench and 8 to 10 on tau2-bench airline.
Full results, figures and method: logits.com/benchmarks/abliterated.
| Weights | Not released. Not served. |
| Base checkpoint | zai-org/GLM-5.3 at commit ba32f8c1, FP8 |
| Input | text only |
| Evaluated | November 7 to 8, 2026 |
Refusals
Share of requests refused, original β abliterated. Temperature 0, low reasoning effort (GLM-5.3 always thinks) unless the row says otherwise. Qwen3Guard-Gen-8B, run locally, judges every answer.
| Test | Answers judged | Original β abliterated |
|---|---|---|
| JailbreakBench harmful | 100 | 93% β 1% |
| HarmBench, no chem-bio | 172 | 93.6% β 1.2% |
| JailbreakBench harmful, Chinese | 100 | 90% β 11% |
| Harmful, maximum effort, finished answers | 30 and 26 | 96.7% β 0% |
| XSTest safe prompts | 250 | 5.2% β 0.4% |
| JailbreakBench benign | 100 | 10% β 1% |
| Off-topic requests declined, when the system prompt says to | 40 | 95% β 77.5% |
| On-topic requests wrongly declined | 20 | 0% β 0% |
The remaining Chinese refusals (11%) are mostly about sexual content and physical harm. The maximum-effort row counts only answers that finished within an 8,448-token budget: all 30 of the original's and 26 of 30 of the abliterated model's.
Capability
Abliterated minus original, in points, paired on identical questions, with a 95% bootstrap interval over questions. openbench 0.5.3, temperature 1.0, high reasoning effort, output capped at 32,768 tokens. The original's score is its first run.
| Benchmark | Items | Original β abliterated | Difference (95% interval) |
|---|---|---|---|
| GPQA Diamond | 198 questions Γ 3 | 83.5 β 82.7 | β0.8 (β3.4, +1.7) |
| MMLU-Pro | 3,000-question fixed subset | 85.3 β 85.1 | β0.1 (β1.1, +0.8) |
| IFBench, prompt strict | 284 prompts | 48.6 β 45.8 | β2.8 (β7.8, +2.1) |
| SimpleQA-Verified | 1,000 questions | 36.2 β 36.8 | +0.6 (β1.7, +3.0) |
| tau2-bench airline | 50 tasks Γ 3 | 82.0 β 74.0 | β8.0 (β16.7, 0.0) |
We read the tau2-bench drop as a real cost. Against the original's second run it is β10.0 (β18.7, β0.7). tau2-bench rewards an agent for refusing what the airline's policy forbids. The edit may weaken those refusals along with the safety ones. A per-task check could neither confirm nor rule this out. The IFBench drop is probably a cost of about 3 points: the original's two runs differ by +0.7. In tau2-bench the model under test also played the customer, so the scores do not compare with public tau2-bench figures. The 32,768-token cap cuts off some long GPQA answers, so these GPQA scores sit below an uncapped run.
One run each at temperature 0, no interval:
| Benchmark | Items | Original β abliterated |
|---|---|---|
| MMLU | 1,000 questions | 90.2 β 89.3 |
| GSM8K | 500 problems | 94.8 β 96.0 |
| IFEval, prompt strict | 541 prompts | 84.7 β 85.2 |
Trade-offs
- System-prompt restrictions hold less often. Told by the system prompt to decline off-topic requests, the abliterated model declines 77.5% of them, against 95% for the original. Neither declines on-topic requests.
- Fewer "I don't know" answers. On SimpleQA-Verified it answers 89% of questions instead of 77%, and gets 41.5% of the ones it answers right instead of 47.1%. Its total correct moves +0.6 points.
- Answers are longer. Answers use 0.9% to 11.9% more output tokens, depending on the benchmark (IFBench +0.9%, SimpleQA +5.8%, MMLU-Pro +7.1%, GPQA +11.9%).
Serving checks
Measured on the abliterated model only.
| Check | Abliterated |
|---|---|
| Tool calls | 10 / 10 correct |
| Needle at 10%, 50% and 90% of a 54k-token prompt | 3 / 3 found |
| Multi-turn recall | works |
| Multi-token prediction, accepted tokens per draft | 3.09 |
The original was measured on a different workload, so its draft acceptance has no like-for-like baseline.
Method
- Serving. vLLM with 8-way tensor parallelism, an FP8 KV cache and 5-token multi-token prediction.
- Data. Harmful prompts and the model's answers to them never went to a third-party API, and we kept only the scores.
- Not measured. HLE, StrongREJECT, SORRY-Bench and a random-direction control.
License
The base checkpoint is released under zai-org/GLM-5.3's license.